Detecting Concept Drift in Financial Time Series Prediction using Symbolic Machine Learning

This paper investigates the use of strategies to enhance an existing machine learning tool, C4.5, to deal with concept drift and non-determinism in a time series domain. Temporal prediction is a difficult problem faced in most human endeavours. While many specialised time series prediction techniques have been developed, these techniques have limitations. Most are restricted to modeling whole series rather than extracting predictive features and are difficult for domain experts to understand. Symbolic machine learning promises to address these limitations. Symbolic machine learning has been very successful on a broad range of complex problems. To date, few attempts have been made to apply symbolic machine learning directly to temporal prediction. This has resulted in systems that cannot explicitly represent temporally ordered examples or handle changing target concepts. Financial prediction is a challenging target domain, which is temporally ordered, has target concepts that change over time and exhibits a high level of non-determinism. Financial markets are considered to be unpredictable by many academics and thus any improvement on chance is interesting. An aim of this study is to demonstrate that machine learning is capable of providing useful predictive strategies in financial prediction. For short term financial prediction a successful prediction rate of 60% is considered the minimum useful to domain experts. Our results imply that machine learning can exceed this target with the use of new techniques. By trading off coverage for accuracy, we were able to minimise the effect of both noise and concept drift. The work reported was in collaboration with and funded by Australian Gilt Securities Limited. Michael Harries was supported by an Australian Postgraduate Award (Industrial).